Enterprise tax risk prediction method and device, computer equipment and storage medium

Abstract

The invention discloses an enterprise tax risk prediction method and device, computer equipment and a storage medium. The method comprises the steps of obtaining to-be-predicted enterprise tax data; performing feature engineering processing on the to-be-predicted enterprise tax data to obtain a processing result; inputting the processing result into a prediction model for risk value prediction to obtain a prediction result; determining a risk level according to the prediction result and a set risk threshold; outputting the risk level; wherein the prediction model is obtained by training a multi-task learning network through a full sample set composed of a risk sample data set and a non-risk sample data set. By implementing the method provided by the invention, the identification capability of the tax risk can be improved, and the prediction accuracy and comprehensiveness can be improved.

Description

Technical Field

[0001] The present invention relates to a tax risk prediction method, and more specifically to an enterprise tax risk prediction method, device, computer equipment and storage medium. Background Art

[0002] Existing technologies mainly rely on traditional financial auditing and manual inspection methods when predicting corporate tax risks. However, these methods have some significant defects, which affect the effectiveness and accuracy of risk assessment. Specifically, traditional methods usually rely mainly on financial data reported by enterprises themselves, which may be inaccurate or incomplete. In addition, these methods often ignore external public information, such as business registration data, judicial litigation records, and industry dynamics. This lack of information integration makes it difficult to fully understand the financial status of enterprises during risk assessment, which may lead to deviations in risk assessment results. The manual audit process is usually time-consuming and costly, and it is difficult to monitor and assess the tax risks of a large number of enterprises in real time. This inefficiency makes it difficult for auditors to detect potential problems in a timely manner, which may result in missing key early warning signals before tax risks appear. Traditional audit methods have limitations in identifying potential tax risks in non-financial data. For example, non-financial information such as related-party transactions and abnormal changes in corporate executives may indicate tax risks, but traditional methods are difficult to effectively mine these data, thereby missing important clues and affecting the accuracy of overall risk assessment. Current risk assessment methods lack the effective application of big data and intelligent algorithms, and cannot automatically extract and identify valuable information from massive data. This technical deficiency limits the ability to identify risks, resulting in the inability to promptly detect and respond to potential tax risks.

[0003] In summary, existing technologies in corporate tax risk prediction have problems such as single data source, low efficiency, insufficient accuracy and lack of technical application. These factors together affect the comprehensiveness and effectiveness of risk assessment.

[0004] Therefore, it is necessary to develop a new method to improve the ability to identify tax risks and enhance the accuracy and comprehensiveness of forecasts. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and provide a method, device, computer equipment and storage medium for predicting corporate tax risks.

[0006] To achieve the above purpose, the present invention adopts the following technical solution: a method for predicting corporate tax risks, comprising: Obtain the corporate tax data to be predicted; Performing feature engineering processing on the enterprise tax data to be predicted to obtain a processing result; Inputting the processing result into the prediction model to perform risk value prediction to obtain a prediction result; Determine the risk level according to the prediction result and the set risk threshold; outputting the risk level; The prediction model is obtained by training a multi-task learning network with a full sample set consisting of a risk sample data set and a non-risk sample data set.

[0007] A further technical solution is: the tax data to be predicted is subjected to feature engineering processing to obtain a processing result, including: Performing continuous feature WOE conversion on the enterprise tax data to be predicted to obtain a conversion result; Performing named entity recognition on the enterprise tax data to be predicted to obtain key information; Setting rules for processing the enterprise tax data to be predicted to obtain processing results; The conversion result, key information and processing result are integrated into a processing result.

[0008] A further technical solution is: performing named entity recognition on the enterprise tax data to be predicted to obtain key information, including: The BiLSTM-CRF-based deep learning model is used to perform named entity recognition on the corporate tax data to be predicted to obtain key information.

[0009] Its further technical solution is: the prediction model includes an input layer, a feature fusion layer, a gated network layer, an expert network layer and an output layer.

[0010] A further technical solution is: the processing result is input into a prediction model to predict the risk value to obtain a prediction result, including: inputting the processing results into a prediction model; Fusing the processing results to obtain a fusion result; According to the fusion result, the gating network layer determines the expert network layer involved in the processing; Processing the fusion result using a determined expert network layer to determine the tax risk type; The tax risk types output by the expert network are integrated and the tax risk prediction probability is calculated through the Softmax layer to obtain the prediction result.

[0011] Its further technical solution is: the gated network layer determines the expert network layer involved in the processing according to the fusion result, including: The gated network layer assigns tasks to corresponding expert network layers according to the characteristics of the fusion results, wherein each expert network layer focuses on a certain type of tax risk.

[0012] The present invention also provides a device for predicting corporate tax risks, comprising: A data acquisition unit, used to acquire the enterprise tax data to be predicted; A feature engineering processing unit, used for performing feature engineering processing on the enterprise tax data to be predicted to obtain a processing result; A prediction unit, used for inputting the processing result into a prediction model to perform risk value prediction to obtain a prediction result; A risk level determination unit, used to determine the risk level according to the prediction result and a set risk threshold; An output unit, used for outputting the risk level; The prediction model is obtained by training a multi-task learning network with a full sample set consisting of a risk sample data set and a non-risk sample data set.

[0013] Its further technical solution is: the feature engineering processing unit includes: A conversion subunit, used for performing continuous feature WOE conversion on the enterprise tax data to be predicted to obtain a conversion result; An identification subunit, used for performing named entity recognition on the enterprise tax data to be predicted to obtain key information; A processing subunit, used for setting rules for processing the enterprise tax data to be predicted to obtain a processing result; The integration subunit is used to integrate the conversion result, key information and processing result into a processing result.

[0014] The present invention further provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor implements the above method when executing the computer program.

[0015] The present invention also provides a storage medium, wherein the storage medium stores a computer program, and the computer program implements the above method when executed by a processor.

[0016] Compared with the prior art, the present invention has the following beneficial effects: after obtaining the tax data of the enterprise to be predicted, the present invention performs feature engineering processing on it to extract key features. Then, the processing results are input into the prediction model for risk value prediction, ensuring that the model has stronger recognition ability through multi-task learning network training of the full sample set. The risk level is judged according to the prediction results and the set risk threshold, and the risk level is finally output. This process not only improves the ability to identify tax risks, but also enhances the accuracy and comprehensiveness of the prediction.

[0017] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0019] Figure 1 A schematic diagram of an application scenario of the enterprise tax risk prediction method provided by an embodiment of the present invention; Figure 2 A schematic diagram of a process for predicting corporate tax risks provided by an embodiment of the present invention; Figure 3 A schematic block diagram of a device for predicting corporate tax risks provided by an embodiment of the present invention; Figure 4 A schematic block diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0022] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0023] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0024] See also Figure 1 and Figure 2 , Figure 1 A schematic diagram of an application scenario of the enterprise tax risk prediction method provided in an embodiment of the present invention. Figure 2 A schematic flow chart of a method for predicting corporate tax risks provided by an embodiment of the present invention. The method for predicting corporate tax risks is applied in a server. The server interacts with the terminal to collect tax data of the enterprise to be predicted, ensuring the integrity and representativeness of the data; performs feature engineering on the data, including continuous feature WOE and named entity recognition, to extract key information; sets rules to process the data to form a multi-dimensional feature set to enrich the input information; inputs the processing results into a multi-task learning prediction model to improve recognition capabilities; uses a gated network layer in the model to select a suitable expert network, and performs targeted processing according to task characteristics; integrates the outputs of each expert network, and uses the Softmax layer to calculate the probability of tax risk prediction; classifies and outputs different risk levels according to the set risk threshold to ensure comprehensiveness; and improves prediction accuracy and model adaptability by continuously updating sample data sets and model training.

[0025] Figure 2 Schematic diagram of the process of the enterprise tax risk prediction method provided by the embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S150.

[0026] S110. Obtain the enterprise tax data to be predicted.

[0027] In this embodiment, the tax data of the enterprise to be predicted generally includes the following aspects: Financial statement data: such as balance sheet, income statement, cash flow statement, etc., reflect the financial status and operating results of the enterprise.

[0028] Tax declaration information: including tax declaration forms, tax amounts, tax types, declaration time, etc. of previous years.

[0029] Transaction records: including detailed transaction information such as sales, purchases, and expenses, to help analyze the compliance and rationality of transactions.

[0030] Industry characteristics: information such as the tax policies of the industry in which the enterprise is located, the average tax burden of the industry, etc., in order to facilitate industry comparison.

[0031] Basic information of the enterprise: such as registered capital, business scope, shareholder structure, management information, etc., which affect the tax risks of the enterprise.

[0032] Historical audit records: Previous audit results and rectification status can help identify potential tax risk points.

[0033] Compliance data: including the company's compliance management measures, internal control system and its implementation.

[0034] External environmental factors: external factors that may affect tax risks, such as economic conditions, policy changes, market dynamics, etc.

[0035] S120: Perform feature engineering processing on the enterprise tax data to be predicted to obtain a processing result.

[0036] In this embodiment, the processing result refers to the comprehensive data output obtained after feature engineering processing of the enterprise tax data to be predicted. Specifically, it includes the following parts: WOE results: Continuous features converted by WOE (Weight of Evidence) can better reflect the relationship between variables and target variables and improve the prediction ability of the model.

[0037] Key information: Useful information extracted through named entity recognition, such as company name, tax terms, transaction type, etc., helps identify important factors affecting tax risks.

[0038] Processing results: Features obtained after processing data based on set rules, such as tax indicators classified or aggregated according to specific standards, can reveal trends and patterns in the data.

[0039] Ultimately, these processing results are integrated together to form a rich feature set, providing high-quality input data for subsequent predictive models, thereby improving the accuracy and effectiveness of tax risk prediction.

[0040] In one embodiment, the above-mentioned step S120 may include steps S121 to S124.

[0041] S121, performing continuous feature WOE conversion on the enterprise tax data to be predicted to obtain a conversion result.

[0042] In this embodiment, the conversion result refers to the result obtained after the data continuous feature is WOE-ized.

[0043] Specifically, WOE (Weight of Evidence) is a method of processing continuous features in order to improve the convergence and predictive ability of the model. By discretizing continuous variables and converting them into WOE values, the distribution of the variables can be transformed into a form that better meets the requirements of the model.

[0044] Discretization: Divide continuous variables (such as years of establishment, registered capital) into N equal parts. This process can reduce the sensitivity of the model to the distribution assumptions of continuous variables.

[0045] WOE calculation: The WOE value is calculated based on the ratio of positive samples to negative samples in each bucket. The formula is: ,in, represents the number of positive samples in the i-th bucket of x, represents the number of negative samples in the i-th bucket of x, Indicates the number of positive samples in the total sample of x, Indicates the number of negative samples in the total sample.

[0046] After WOE, the x variable is transformed into n feature variables ,use Substitute x into the model.

[0047] After the conversion, the interpretability of the model is improved and the complexity of the model is reduced; it can better handle the problem of sample imbalance; it enhances the convergence of the model and is suitable for the input of deep learning models.

[0048] S122. Perform named entity recognition on the enterprise tax data to be predicted to obtain key information.

[0049] Specifically, a deep learning model based on BiLSTM-CRF is used to perform named entity recognition on the corporate tax data to be predicted to obtain key information.

[0050] In this embodiment, named entity recognition (NER) technology is used to extract key business terms from the text content of the enterprise name to better identify the business scope, industry category and potential risks of the enterprise. NER technology is usually used to identify named entities in text, such as names of people, institutions and places. In the present invention, NER is introduced to automatically identify industry-related keywords in enterprise names, such as industry type, geographical identification and business scope, etc. These keywords help to improve the accuracy and comprehensiveness of the enterprise risk prediction model.

[0051] The NER recognition module can automatically extract the following key information from company names: Industry keywords: such as "technology", "trade", "manufacturing", "investment", etc., to identify the industry the company is involved in.

[0052] Geographical identification: For example, "City A", "City B", "Province C", etc., to identify the geographical location or place of registration of the enterprise.

[0053] Business type: For example, "Import and Export", "E-commerce", "Consulting Services", etc., to identify the business scope and direction of the enterprise.

[0054] This module performs NER based on the BiLSTM-CRF deep learning model. Its core steps are as follows: Data preprocessing: Perform basic processing on company names, including removing special characters, word segmentation, and unifying upper and lower case to ensure consistency in model processing.

[0055] Model training: The NER model is trained using a dataset of company names containing annotated entities. The annotated entity types include "industry category", "regional identifier" and "business type". Combining BiLSTM (bidirectional long short-term memory network) and CRF (conditional random field) technology, the model can effectively identify keyword entities in company names.

[0056] Keyword recognition and extraction: The trained NER model automatically recognizes key information in company names. For example, "A City Technology Co., Ltd." is recognized as {region: A City, industry: technology}, while "B City E-commerce Group" is recognized as {region: B City, business type: e-commerce}.

[0057] The identified keywords will be standardized to ensure their uniqueness and operability in the system. Each keyword will be assigned a unique code for subsequent processing and matching. For example: “Industry Category: Technology” might be coded as {10:1001}; "Region: City A" may be coded as {20:2001}; "Business Type: E-commerce" might be coded as {30:3001}.

[0058] This standardization not only ensures the consistency of various keywords in the system, but also improves the indexing and query efficiency during data processing. Automated key information extraction improves the efficiency of data processing. It provides a more accurate basis for enterprise classification and risk assessment. It provides structured information support for subsequent data analysis and model building.

[0059] S123, setting rules for processing the enterprise tax data to be predicted to obtain a processing result.

[0060] In this embodiment, at this stage, multiple features are identified and processed, including basic information features, industry features, shareholder and executive features, business features, change record features, and blacklist features.

[0061] Basic information characteristics: such as unified social credit code, company name, registered capital, etc. These characteristics are used to describe the basic situation of the company.

[0062] Industry characteristics: The industry code to which the enterprise belongs can help analyze the industry background and market competition.

[0063] Shareholder and executive characteristics: Information such as shareholder stability and the number of executives helps assess a company’s governance structure and management level.

[0064] Business characteristics: Characteristics such as incorporation date and registered capital can help determine the financial health of a business.

[0065] Change record feature: records changes in the enterprise and provides a basis for dynamic risk assessment.

[0066] Blacklist features: By checking whether the enterprise, legal person, shareholder, and executive are on the blacklist, it helps to quickly identify potential high-risk enterprises.

[0067] Industry average characteristics: Provides average level data within the industry to help relatively evaluate the performance of enterprises.

[0068] Through multi-dimensional feature construction, the accuracy and comprehensiveness of the risk assessment model are improved. It is convenient for subsequent model training and prediction, and enhances the robustness of the model. It provides a systematic information framework to facilitate subsequent decision-making analysis.

[0069] In addition, the setting rules for detailed processing of features are shown in Tables 1 to 7.

[0070] Table 1 Field English name Field Chinese name describe taxpayer_id Unified social credit code org_encode Organization Code company_name Company Name After NER processing, it is fed into the self-developed deep learning network. register_capital Registered capital register_province Registered Province register_city Registered City operating_status Operating Status Table 2 ent_type Company Type ent_type_stat Enterprise type (after mapping) taxpayer_industry1 Authorized enterprise level industry taxpayer_industry2 Authorized enterprise secondary industry taxpayer_industry3 Authorized enterprise level 3 industry taxpayer_industry4 Authorized Enterprise Level 4 Industry shareholders_stability Shareholder stability number_managers Number of managers Table 3

[0071] Table 4

[0072] Table 5

[0073] Table 6

[0074] Table 7

[0075] S124, integrating the conversion result, key information and processing result into a processing result.

[0076] The conversion result, key information and processing result are combined into a processing result.

[0077] S130, inputting the processing result into a prediction model to perform risk value prediction to obtain a prediction result.

[0078] In this embodiment, the prediction result refers to the probability value of the enterprise tax risk belonging to the corresponding type.

[0079] The prediction model is obtained by training a multi-task learning network with a full sample set consisting of a risk sample data set and a non-risk sample data set.

[0080] The prediction model includes an input layer, a feature fusion layer, a gating network layer, an expert network layer and an output layer.

[0081] The training process of the prediction model is as follows: First, collect public data, such as the list of companies involved in major tax violations published by the tax authorities; regularly (T+1 update) collect the list of companies involved in major tax violations published by the tax authorities, and organize the data according to two types: tax risk and risk description.

[0082] In addition, the company collects real tax violation case data accumulated over a long period of time. This data is only used in the model training phase and is not used in the model inference phase.

[0083] The list of Class A taxpayers published by the tax authorities is collected regularly (updated on T+1) to obtain the aggregate LA.

[0084] After combining the public data with the real tax violation case data accumulated by the company over a long period of time, the company list obtained includes the registered addresses of the companies forming the blacklist address library, recorded as set H1; the legal persons of the companies forming the blacklist legal person library, recorded as set H2; the shareholders of the companies forming the blacklist shareholder library, recorded as set H3; and the senior executives of the companies forming the blacklist senior executive library, recorded as set H4.

[0085] Next, we merged the public data with the real tax violation case data accumulated by the company over a long period of time to form data set A1. Then we screened the enterprises with the operating status of "existing" or "in operation" to obtain data set A2. If the same enterprise has multiple violations, they will be recorded as multiple training samples, but we must ensure that the same enterprise does not appear in the training set, test set, and validation set at the same time. Bad samples are classified according to the 10 types in Tables 1 to 7. If the enterprise hits any type, it can be determined that it has tax risks.

[0086] From all enterprises in the "existing" or "in operation" status, enterprises in data set A1 are eliminated to form data set B1. Then, stratified sampling is performed based on dimensions such as enterprise size, industry, and region, and good samples are selected from data set B1 to ensure the representativeness and balance of the samples, and finally data set B2 is obtained.

[0087] Merge dataset A2 and dataset B2 to get the full sample set C1. During the training process, split C1 into training set, test set, and validation set to ensure that the same company does not appear in multiple datasets at the same time. Finally, the full sample set C1 is the union of dataset A2 and dataset B2.

[0088] The full sample set C1 is subjected to feature engineering, as described in step S120, which will not be repeated here. To ensure data security and model accuracy, all features are taken from the last statistical period before the tax risk occurs. A feature extraction process is designed to ensure that each piece of data can be automatically obtained and processed from public resources. This includes natural language processing technology to parse complex text data, using xgbooxt to evaluate and select the importance of features, filling missing values, detecting outliers, and normalizing data.

[0089] Build a multi-task learning network. The multi-task learning network is based on the application of the Mixture of Experts (MoE) architecture in the multi-task learning model, which effectively handles various tax risk prediction scenarios. It includes the input layer, feature fusion layer, gating network layer, expert network layer and output layer, from the input layer through feature fusion and gating and processing with the expert network, and finally to the complete process of the output layer. This figure reflects the core architecture of the MoE model, including the allocation of expert networks and the dynamic adjustment of gating weights; the processing tasks are dynamically allocated to different expert networks through the gating mechanism of the MoE architecture. Each expert network handles a specific tax risk type, and then merges its output into the fusion layer. Then, the various tax risk predictions from the expert network are integrated into a comprehensive output, and finally processed through the Softmax layer to output the overall tax risk probability.

[0090] The next step is model training. The model training step is a key step, which involves using the prepared full sample set C1 (including the bad sample data set A2 and the good sample data set B2) to train a multi-task learning model based on the Mixture of Experts (MoE) architecture. The following are the detailed working steps: Data preprocessing: Perform data cleaning on the full sample set C1, including processing missing values, detecting outliers, and standardizing data to ensure that the data quality is suitable for model training.

[0091] Feature selection and conversion: Based on the results of feature engineering, select features that have strong predictive power for predicting corporate tax risks. At the same time, convert text data into a format that can be processed by the model through natural language processing technology.

[0092] Sample segmentation: The full sample set C1 is randomly divided into a training set, a validation set, and a test set. The training set is used to learn the model, the validation set is used to adjust the model parameters to prevent overfitting, and the test set is used to evaluate the generalization ability of the model. Ensure that the same enterprise does not appear in different data sets at the same time.

[0093] Model architecture settings: Configure the Mixture of Experts architecture, design multiple expert networks and a gating mechanism. Each expert network focuses on learning the characteristic patterns of a specific type of tax risk, while the gating mechanism is responsible for dynamically selecting the most suitable expert network based on the input features.

[0094] Model training: Use the gradient descent algorithm to optimize model parameters. During the training process, continuously monitor the performance of the validation set and apply techniques such as early stopping to avoid overfitting. At the same time, methods such as weight decay and dropout can be used to improve the generalization ability of the model.

[0095] Hyperparameter tuning: Use automated hyperparameter optimization techniques (such as grid search and Bayesian optimization) to find the best model parameters, including learning rate, batch size, and number of experts.

[0096] Model evaluation and selection: Evaluate model performance on the test set, paying special attention to the prediction accuracy and recall of different types of tax risks. Finally, select the model configuration that performs best on multiple indicators.

[0097] In one embodiment, the above-mentioned step S130 may include steps S131 to S135.

[0098] S131, inputting the processing result into the prediction model.

[0099] In this stage, the processing results obtained in the previous step (which may be feature-extracted data or preprocessing results) are first passed to the prediction model. The prediction model can be a machine learning or deep learning-based model for further analysis and prediction.

[0100] Inputting cleaned and characterized data directly into the model can improve the accuracy and efficiency of subsequent predictions, forming a systematic data processing chain, reducing the need for manual intervention, and thus improving consistency.

[0101] S132: Fusing the processing results to obtain a fusion result.

[0102] In this embodiment, data from different sources or with different characteristics are fused, and this step is usually achieved through weighted averaging, splicing or other fusion technologies.

[0103] Fusion of different information sources can provide a more comprehensive perspective and reduce the one-sidedness of information. Fusion of data can improve the robustness and generalization ability of the model and reduce the noise impact that a single data source may bring.

[0104] S133. The gated network layer determines the expert network layer involved in the processing according to the fusion result.

[0105] Specifically, the gated network layer assigns tasks to corresponding expert network layers according to the characteristics of the fusion result, wherein each expert network layer focuses on a certain type of tax risk.

[0106] The gating network layer dynamically decides which expert networks will participate in the final prediction based on the characteristics of the fusion results. This process includes evaluating the suitability of each expert and assigning the task to the most relevant expert network.

[0107] Each expert network is trained for a specific type of tax risk, so that the model can handle different situations in a targeted manner and improve the accuracy of predictions. By selecting the most suitable expert network, the waste of computing resources is reduced and the processing speed is improved.

[0108] S134. Process the fusion result using the determined expert network layer to determine the tax risk type.

[0109] In this embodiment, the expert network selected by the gated network layer will process the fusion result and output the type of tax risk. This step usually involves the use of forward propagation and activation function.

[0110] By focusing on specific types of tasks, expert networks are able to identify and classify tax risks faster. The specialized training of each expert network makes it perform better on a specific task, further improving the accuracy of predictions.

[0111] S135. Integrate the tax risk types output by the expert network, and calculate the tax risk prediction probability through the Softmax layer to obtain the prediction result.

[0112] In this embodiment, the tax risk types output by all expert networks will eventually be integrated, and then the predicted probability of each tax risk type will be calculated through the Softmax layer. This process ensures the interpretability of the final result.

[0113] Using the Softmax layer can provide the predicted probability of each tax risk, making the decision-making process more transparent and explainable. The probabilistic results can help decision makers understand the relative likelihood of different risks, so as to better manage risks and allocate resources.

[0114] Through the implementation of the above steps, the entire forecasting process can efficiently and accurately identify and assess the tax risks of enterprises and improve the decision-making support capabilities.

[0115] S140: Determine the risk level according to the prediction result and the set risk threshold.

[0116] In this embodiment, the tax risk of the enterprise is evaluated according to the set threshold. Usually, the evaluation result will give both the risk probability value and the risk level. The specific classification is as follows: A-level risk: score above 90 points; Level B risk: score between 70 and 90 points; C-level risk: score between 50 and 70 points; D-level risk: score below 50 points; These thresholds can be adjusted dynamically based on actual business needs.

[0117] S150: Output the risk level.

[0118] In this embodiment, the risk level is sent to the terminal for display.

[0119] The above-mentioned enterprise tax risk prediction method obtains the tax data of the enterprise to be predicted, and then performs feature engineering processing on it to extract key features. Then, the processing results are input into the prediction model for risk value prediction to ensure that the model has stronger recognition ability through multi-task learning network training of the full sample set. The risk level is judged according to the prediction results and the set risk threshold, and the risk level is finally output. This process not only improves the ability to identify tax risks, but also enhances the accuracy and comprehensiveness of the prediction.

[0120] Figure 3 is a schematic block diagram of an enterprise tax risk prediction device 300 provided in an embodiment of the present invention. Figure 3 As shown, corresponding to the above enterprise tax risk prediction method, the present invention also provides an enterprise tax risk prediction device 300. The enterprise tax risk prediction device 300 includes a unit for executing the above enterprise tax risk prediction method, and the device can be configured in a server. Figure 3 The enterprise tax risk prediction device 300 includes a data acquisition unit 301, a feature engineering processing unit 302, a prediction unit 303, a risk level determination unit 304 and an output unit 305.

[0121] A data acquisition unit 301 is used to acquire the enterprise tax data to be predicted; a feature engineering processing unit 302 is used to perform feature engineering processing on the enterprise tax data to be predicted to obtain a processing result; a prediction unit 303 is used to input the processing result into a prediction model to perform risk value prediction to obtain a prediction result; a risk level determination unit 304 is used to determine the risk level according to the prediction result and a set risk threshold; an output unit 305 is used to output the risk level; wherein the prediction model is obtained by training a multi-task learning network with a full sample set consisting of a risk sample data set and a non-risk sample data set.

[0122] In one embodiment, the feature engineering processing unit 302 includes: The conversion subunit is used to perform continuous feature WOE on the corporate tax data to be predicted to obtain a conversion result; the identification subunit is used to perform named entity recognition on the corporate tax data to be predicted to obtain key information; the processing subunit is used to set rules for processing the corporate tax data to be predicted to obtain a processing result; the integration subunit is used to integrate the conversion result, key information and processing result into a processing result.

[0123] In one embodiment, the prediction unit 303 includes: An input subunit is used to input the processing result into the prediction model; a fusion subunit is used to fuse the processing results to obtain a fusion result; a determination subunit is used to determine the expert network layer involved in the processing according to the fusion result by the gated network layer; a processing subunit is used to process the fusion result using the determined expert network layer to determine the tax risk type; an integration subunit is used to integrate the tax risk types output by the expert network, and calculate the tax risk prediction probability through the Softmax layer to obtain the prediction result.

[0124] In one embodiment, the determination subunit is used for the gating network layer to assign tasks to corresponding expert network layers according to the characteristics of the fusion result, wherein each expert network layer focuses on a certain type of tax risk.

[0125] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned enterprise tax risk prediction device 300 and each unit can refer to the corresponding description in the aforementioned method embodiment, and for the convenience and brevity of description, it will not be repeated here.

[0126] The enterprise tax risk prediction device 300 can be implemented in the form of a computer program. The computer program can be used in Figure 4 Runs on the computer device shown.

[0127] See also Figure 4 , Figure 4 5 is a schematic block diagram of a computer device provided in an embodiment of the present application. The computer device 500 may be a server, wherein the server may be an independent server or a server cluster composed of multiple servers.

[0128] See also Figure 4 The computer device 500 includes a processor 502 , a memory and a network interface 505 connected via a system bus 501 , wherein the memory may include a non-volatile storage medium 503 and an internal memory 504 .

[0129] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions, and when the program instructions are executed, the processor 502 can execute a method for predicting corporate tax risks.

[0130] The processor 502 is used to provide computing and control capabilities to support the operation of the entire computer device 500 .

[0131] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for predicting corporate tax risks.

[0132] The network interface 505 is used to communicate with other devices over the network. Figure 4 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application, and does not constitute a limitation on the computer device 500 to which the solution of the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0133] The processor 502 is used to run the computer program 5032 stored in the memory to implement the following steps: Acquire the enterprise tax data to be predicted; perform feature engineering processing on the enterprise tax data to be predicted to obtain a processing result; input the processing result into the prediction model to perform risk value prediction to obtain a prediction result; determine the risk level according to the prediction result and the set risk threshold; output the risk level; The prediction model is obtained by training a multi-task learning network with a full sample set consisting of a risk sample data set and a non-risk sample data set.

[0134] The prediction model includes an input layer, a feature fusion layer, a gating network layer, an expert network layer and an output layer.

[0135] In one embodiment, when the processor 502 performs feature engineering processing on the enterprise tax data to be predicted to obtain a processing result, the processor 502 specifically implements the following steps: The enterprise tax data to be predicted is subjected to continuous feature WOE conversion to obtain a conversion result; the enterprise tax data to be predicted is subjected to named entity recognition to obtain key information; the enterprise tax data to be predicted is subjected to set rule processing to obtain a processing result; the conversion result, key information and processing result are integrated into a processing result.

[0136] In one embodiment, when the processor 502 implements the step of performing named entity recognition on the enterprise tax data to be predicted to obtain key information, the processor 502 specifically implements the following steps: The BiLSTM-CRF-based deep learning model is used to perform named entity recognition on the corporate tax data to be predicted to obtain key information.

[0137] In one embodiment, when the processor 502 implements the step of inputting the processing result into the prediction model to perform risk value prediction to obtain the prediction result, the processor 502 specifically implements the following steps: The processing results are input into the prediction model; the processing results are fused to obtain a fusion result; the gated network layer determines the expert network layer involved in the processing according to the fusion result; the fusion result is processed by the determined expert network layer to determine the tax risk type; the tax risk types output by the expert network are integrated, and the tax risk prediction probability is calculated through the Softmax layer to obtain the prediction result.

[0138] In one embodiment, when the processor 502 implements the step of determining, by the gated network layer according to the fusion result, the expert network layer involved in the processing, the processor 502 specifically implements the following steps: The gated network layer assigns tasks to corresponding expert network layers according to the characteristics of the fusion results, wherein each expert network layer focuses on a certain type of tax risk.

[0139] It should be understood that in the embodiment of the present application, the processor 502 may be a central processing unit (CPU), and the processor 502 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0140] It can be understood by those skilled in the art that all or part of the processes in the method for implementing the above embodiment can be completed by instructing the relevant hardware through a computer program. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiment of the above method.

[0141] Therefore, the present invention also provides a storage medium. The storage medium may be a computer-readable storage medium. The storage medium stores a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps: Acquire the enterprise tax data to be predicted; perform feature engineering processing on the enterprise tax data to be predicted to obtain a processing result; input the processing result into the prediction model to perform risk value prediction to obtain a prediction result; determine the risk level according to the prediction result and the set risk threshold; output the risk level; The prediction model is obtained by training a multi-task learning network with a full sample set consisting of a risk sample data set and a non-risk sample data set.

[0142] The prediction model includes an input layer, a feature fusion layer, a gating network layer, an expert network layer and an output layer.

[0143] In one embodiment, when the processor executes the computer program to implement the step of performing feature engineering processing on the enterprise tax data to be predicted to obtain a processing result, the processor specifically implements the following steps: The enterprise tax data to be predicted is subjected to continuous feature WOE conversion to obtain a conversion result; the enterprise tax data to be predicted is subjected to named entity recognition to obtain key information; the enterprise tax data to be predicted is subjected to set rule processing to obtain a processing result; the conversion result, key information and processing result are integrated into a processing result.

[0144] In one embodiment, when the processor executes the computer program to implement the step of performing named entity recognition on the enterprise tax data to be predicted to obtain key information, the processor specifically implements the following steps: The BiLSTM-CRF-based deep learning model is used to perform named entity recognition on the corporate tax data to be predicted to obtain key information.

[0145] In one embodiment, when the processor executes the computer program to implement the step of inputting the processing result into the prediction model to perform risk value prediction to obtain the prediction result, the processor specifically implements the following steps: The processing results are input into the prediction model; the processing results are fused to obtain a fusion result; the gated network layer determines the expert network layer involved in the processing according to the fusion result; the fusion result is processed by the determined expert network layer to determine the tax risk type; the tax risk types output by the expert network are integrated, and the tax risk prediction probability is calculated through the Softmax layer to obtain the prediction result.

[0146] In one embodiment, when the processor executes the computer program to implement the step of determining, by the gated network layer according to the fusion result, the expert network layer involved in the processing, the processor specifically implements the following steps: The gated network layer assigns tasks to corresponding expert network layers according to the characteristics of the fusion results, wherein each expert network layer focuses on a certain type of tax risk.

[0147] The storage medium may be a USB flash drive, a mobile hard disk, a read-only memory (ROM), a magnetic disk, or an optical disk, etc., which are computer-readable storage media that can store program codes.

[0148] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0149] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of each unit is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0150] The steps in the method of the embodiment of the present invention can be adjusted in order, combined and deleted according to actual needs. The units in the device of the embodiment of the present invention can be combined, divided and deleted according to actual needs. In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0151] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, terminal, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention.

[0152] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. The enterprise tax risk prediction method is characterized by: include: Obtain the corporate tax data to be predicted; Performing feature engineering processing on the enterprise tax data to be predicted to obtain a processing result; Inputting the processing result into the prediction model to perform risk value prediction to obtain a prediction result; Determine the risk level according to the prediction result and the set risk threshold; outputting the risk level; Wherein, the prediction model is obtained by training a multi-task learning network with a full sample set consisting of a risk sample data set and a non-risk sample data set; The step of performing feature engineering processing on the enterprise tax data to be predicted to obtain a processing result includes: Performing continuous feature WOE conversion on the enterprise tax data to be predicted to obtain a conversion result; Performing named entity recognition on the enterprise tax data to be predicted to obtain key information; Setting rules for processing the enterprise tax data to be predicted to obtain processing results; The conversion result, key information and processing result are integrated into a processing result.

2. The enterprise tax risk prediction method according to claim 1 is characterized in that: The named entity recognition is performed on the enterprise tax data to be predicted to obtain key information, including: The BiLSTM-CRF-based deep learning model is used to perform named entity recognition on the corporate tax data to be predicted to obtain key information.

3. The enterprise tax risk prediction method according to claim 1 is characterized in that: The prediction model includes an input layer, a feature fusion layer, a gating network layer, an expert network layer and an output layer.

4. The enterprise tax risk prediction method according to claim 3 is characterized in that: The step of inputting the processing result into a prediction model to perform risk value prediction to obtain a prediction result includes: inputting the processing results into a prediction model; Fusing the processing results to obtain a fusion result; According to the fusion result, the gating network layer determines the expert network layer involved in the processing; Processing the fusion result using a determined expert network layer to determine the tax risk type; The tax risk types output by the expert network are integrated, and the tax risk prediction probability is calculated through the Softmax layer to obtain the prediction result.

5. The enterprise tax risk prediction method according to claim 4 is characterized in that: The gated network layer determines the expert network layer involved in the processing according to the fusion result, including: The gated network layer assigns tasks to corresponding expert network layers according to the characteristics of the fusion results, wherein each expert network layer focuses on a certain type of tax risk.

6. Enterprise tax risk prediction device, characterized in that: include: A data acquisition unit, used to acquire the enterprise tax data to be predicted; A feature engineering processing unit, used for performing feature engineering processing on the enterprise tax data to be predicted to obtain a processing result; A prediction unit, used for inputting the processing result into a prediction model to perform risk value prediction to obtain a prediction result; A risk level determination unit, used to determine the risk level according to the prediction result and a set risk threshold; An output unit, used for outputting the risk level; Wherein, the prediction model is obtained by training a multi-task learning network with a full sample set consisting of a risk sample data set and a non-risk sample data set; The feature engineering processing unit includes: A conversion subunit, used for performing continuous feature WOE conversion on the enterprise tax data to be predicted to obtain a conversion result; An identification subunit, used for performing named entity recognition on the enterprise tax data to be predicted to obtain key information; A processing subunit, used for setting rules for processing the enterprise tax data to be predicted to obtain a processing result; The integration subunit is used to integrate the conversion result, key information and processing result into a processing result.

7. A computer device, characterized in that: The computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 5 when executing the computer program.

8. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images